Guide
How to verify a record before AI training
Updated September 14, 2026 · Maintained by SuperTruth, Inc.
To verify a record before it enters an AI training set, run ten checks in order: that the file arrived intact, where the record came from, whether consent covers training, how current it is for its kind, how complete it is, whether it agrees with independent sources, whether its data type is validated against outcomes, how much it covers, how stable its measurements are, and whether it clears the floor you set. Each check below is one the Data Trust Index (DTI), SuperTruth's patented 0 to 100 score, performs on every record at intake; the weights shown are the published defaults, calibrated in health. A record that fails the floor is held, not trained on.
- 01
Confirm the file arrived intact
Before you ask anything of a record, prove the file you hold is the file that was sent. At the DTI Engine's first stage, Receive, every file is transferred over TLS 1.3, its SHA-256 hash is verified on receipt, and the original is stored and never modified. A record that fails this check is not a record yet; it is a question about transport.
Check: Not a scored dimension: intake integrity at stage 1 of the pipe.
DTI Engine, File ingestion (technical architecture), supertruth.ai/products/dti-engine, read 14 Sep 2026.
- 02
Establish where the record came from
Score the source, not the format. Source pedigree (a CLIA or CAP certified laboratory outranks a value retyped from a fax), device attestation (did the instrument sign the reading), and chain-of-custody documentation (every hand between capture and you). A record that cannot say where it came from is marked down harder than for any other single failure. The data provenance entry has the full definition.
Dimension: Provenance, 25 percent of the default weight
Weight: DTI white paper, DOI 10.5281/zenodo.19601616, 2026. Measures: supertruth.ai/products/dti-engine, read 14 Sep 2026.
- 03
Check that consent covers this use
Read the consent as a property of the record: how explicit it was, whether its scope matches training, how long it runs, and whether it has been revoked since. A record consented for care is not thereby consented for training; the scope has to say so. Which law governs a given record is a question for your counsel; the check here is whether the record can show the permission it claims.
Dimension: Consent, 20 percent of the default weight
Weight: DTI white paper, DOI 10.5281/zenodo.19601616, 2026. Measures: supertruth.ai/products/dti-engine, read 14 Sep 2026.
- 04
Check the record is current for its kind
Age the value against a decay curve for its modality and read the last-updated timestamp. A CBC result ages differently than a continuous HRV stream, and a medication list from 2019 can look perfectly clean and be wrong today. Data does not announce that it has expired; the temporal drift entry describes the failure.
Dimension: Recency, 15 percent of the default weight
Weight: DTI white paper, DOI 10.5281/zenodo.19601616, 2026. Measures: supertruth.ai/products/dti-engine, read 14 Sep 2026.
- 05
Check the record itself
This is the check most teams already run, and it is one of eight. Field completeness, formatting, noise floor and missingness patterns. A field that suddenly empties or changes shape loses points at once; the schema drift entry describes how.
Dimension: Quality, 10 percent of the default weight
Weight: DTI white paper, DOI 10.5281/zenodo.19601616, 2026. Measures: supertruth.ai/products/dti-engine, read 14 Sep 2026.
- 06
Check it agrees with independent sources
Corroborate across sources that did not copy from each other: does the lab result agree with the wearable, does the chart match the claim, is the value consistent over time. Disagreement is a warning, not something to average away; a record whose sources conflict goes to a human before it goes to a model.
Dimension: Concordance, 10 percent of the default weight
Weight: DTI white paper, DOI 10.5281/zenodo.19601616, 2026. Measures: supertruth.ai/products/dti-engine, read 14 Sep 2026.
- 07
Check the data type is validated
Ask whether this kind of measurement is established as meaningful for the decision it will inform: outcome linkage, peer-review status, clinical evidence grade. A novel measure with no outcome linkage scores lower here however clean the value.
Dimension: Validation, 10 percent of the default weight
Weight: DTI white paper, DOI 10.5281/zenodo.19601616, 2026. Measures: supertruth.ai/products/dti-engine, read 14 Sep 2026.
- 08
Check how much the record covers
Count what the record holds: biomarker count, biometric coverage, contextual signals. A richer record is a more trustworthy one, because each additional independent field is another chance to catch a value that does not fit.
Dimension: Breadth, 5 percent of the default weight
Weight: DTI white paper, DOI 10.5281/zenodo.19601616, 2026. Measures: supertruth.ai/products/dti-engine, read 14 Sep 2026.
- 09
Check the measurements are stable
Compare repeated measurements and read the test-retest reliability. A source whose measurements stop agreeing with their own history is marked down before a model reads them.
Dimension: Stability, 5 percent of the default weight
Weight: DTI white paper, DOI 10.5281/zenodo.19601616, 2026. Measures: supertruth.ai/products/dti-engine, read 14 Sep 2026.
- 10
Set the floor and hold what fails
Combine the eight weighted dimensions into one 0 to 100 score, then set a floor per use. Nothing below the floor reaches the model. Under the default tiers a record below 55 is held for remediation and does not enter any tiered use; Bronze (55 to 69) is exploratory, Silver (70 to 79) operational-grade, Gold (80 to 89) clinical-grade and Platinum (90 and above) regulatory-grade. A training set for a clinical model might admit Gold and above; a discovery cohort might admit Silver.
Check: The composite and the floor
Tiers and fail band: DTI white paper, DOI 10.5281/zenodo.19601616, 2026. Floor per use: DTI Engine, Where scored records land, supertruth.ai/products/dti-engine, read 14 Sep 2026.
Seal the score with the record
The score is computed at ingestion and sealed with the record. Every pipeline event is written to an append-only ledger, a cryptographic signature covers the record identifier, the score, the timestamp and the flags, and any field modification breaks the signature. Scoring is deterministic, with no model inference, so an auditor can replay the score from the inputs and the weights and get the same number (DTI Engine, Audit trail and DTI scoring engine, technical architecture, read 14 Sep 2026).
Re-check when anything changes
Scoring runs at intake, at query time and continuously in the background. A source that degrades over time degrades in score. A consent revocation re-scores the record everywhere it is used, so a record that was in a training set last quarter can be out of the next one (DTI Engine, Scoring at every stage; the consent management entry).
Try the ten checks on a synthetic record
Run the pipeline yourself on the DTI Engine page: three synthetic records, none describing a real person, scored 90, 85 and 65. The place beside the record is DataSpine, the sourced record of every US place with source and vintage on every value. The agent that trains on the record is scored while it works by VIGIL. SuperTruth does not use patient data to train models (DTI Engine, Security and compliance).
Sources
- Weights, tiers and the fail band: "The Data Trust Index: a scoring framework for health data trust," SuperTruth Inc., 2026, DOI 10.5281/zenodo.19601616. Cite the DOI rather than this website for the methodology.
- Dimension descriptors and measures, the pipeline stages and the technical architecture: DTI Engine, read 14 September 2026.
- Definitions: Data Trust Index, data provenance, data truth.